New DVFace model uses single-step diffusion with dual spatial-temporal priors to restore degraded video faces faster and more realistically than existing methods.

DVFace introduces a one-step diffusion framework designed specifically for real-world video face restoration, improving both speed and efficiency compared to multi-step approaches

The model employs a spatio-temporal dual-codebook design that extracts complementary spatial and temporal facial priors from degraded videos for better facial adaptation

Uses asymmetric spatio-temporal fusion to maintain realistic facial details, stable identity, and temporal coherence across video frames

Addresses limitations of generic diffusion priors by incorporating face-specific priors, enabling both faithful facial recovery and temporally stable outputs